Dr Rudy Coquet 1 High speed neural network for chemicals and materials discovery Topic: #digital: AI and data analytics Session: AI-based modelling and engineering June 11, 2024

© Preferred Computational Chemistry, Inc. All Rights Reserved. Agenda 2 ❏ Introduction to our company and product (10’) ❏ Use cases with real-world applications (15’) ❏ Catalysts: Renewable synthetic fuels ❏ Battery materials: Lithium diffusion in solid electrolytes ❏ Lubricants: Base oils in silico screening (Dr Tasuku Onodera, ENEOS Corp.) ❏ Q&A (5’)

© Preferred Computational Chemistry, Inc. All Rights Reserved. > Established June 2021 in Tokyo, “to accelerate materials discovery for a sustainable future”, providing Matlantis™, a high-speed universal atomistic simulator. 3 About Us Largest Japanese oil company. https://www.eneos.co.jp/english/ Japan’s AI technology leader. https://www.preferred.jp/en/　 Additionally, since June 2024 :

© Preferred Computational Chemistry, Inc. All Rights Reserved. 4 Parent companies Largest Japanese oil company Japan’s AI technology leader - Part of ENEOS Holdings Group - Capital: 30 billion JPY (~191M €) - Operating income: 82 billion JPY in FY2020 (523M €) - 9,348 employees - Founded in March 2014 - Today the largest unlisted AI tech venture in Japan - R&D business with a wide variety of companies - Developed its own processors, called “MN-Core” - Supercomputer ranked #1 in the Green500 list several times PFN supercomputing cluster https://www.preferred.jp/en/　 https://www.eneos.co.jp/english/

© Preferred Computational Chemistry, Inc. All Rights Reserved. ● Used by 80 organizations (https://matlantis.com/) ● 500 licensed users ● Industries: Academia, Chemicals, Electronics, Mining, Rubber, Ceramics, Automobiles, Non-ferrous metals, Petroleum, etc. ● First academic & commercial organizations outside Japan 5 Business status May 2022 Nov. 2022 2024 2023 July 2021 EU servers From where Matlantis is available US servers Where Matlantis calculations run JP servers …

© Preferred Computational Chemistry, Inc. All Rights Reserved. Matlantis ™: a paradigm shift Lin and Wang, Commun Mater 4, 66 (2023) Going from a few hundred atoms in DFT to 10,000 atoms in Matlantis™, one can study realistic systems, opening the door for exploration into a myriad of materials, chemical, and even biological systems. 6

© Preferred Computational Chemistry, Inc. All Rights Reserved. Three key features 96 elements 7

© Preferred Computational Chemistry, Inc. All Rights Reserved. Accurate - Computes energies and forces from atomic structures with DFT accuracy (cf. use cases) Scalable and secure - No need to install hardware or software. Global standard cloud system (AWS) Flexible and customizable - Programmable environment in JupyterLab (Python) Physical properties calculation library User interface Capabilities / 8

© Preferred Computational Chemistry, Inc. All Rights Reserved. Jupyter Notebook Department A Inference System Group Drive Department B Batch Inference GPUs (Tenant ABC, Inc.) *System configurations and specifications are subject to change ABC, Inc. [ Independent for each user ] - # of instances provided = # of active users - No access rights to other users' data - File system independent for each user - 2 CPU, Memory: 8 GB, Disk: 100 GB / user [ Tokens are allocated to each tenant ] - Used for inference of energy, force and charge. - The amount of Tokens varies depending on the contracted plan - If a tenant's usage exceeds the Token limit, the calculation speed will be limited - Optimized total throughput by batch inference system [ Independent for each tenant ] - Accessible by all users in your tenant - Not visible to other tenants - Size may vary by the purchased license System environment 9 Token: function of the number of atoms that can be processed. Given every second.

© Preferred Computational Chemistry, Inc. All Rights Reserved. Mechanism behind Matlantis ™ PFN cluster with own MN-Core™ series of deep-learning processors 10

© Preferred Computational Chemistry, Inc. All Rights Reserved. Proprietary & Unique Neural Network Potential: TeaNet/PFP Research papers on our NNP: https://www.nature.com/articles/s41467-022-30687-9 https://doi.org/10.1016/j.jmat.2022.12.007 > Our training dataset is generated by molecular dynamics simulation using the machine learning model itself, rather than humans, thereby minimizing the influence of assumed knowledge. > To achieve universality, our dataset includes disordered and metastable structures: this is a unique feature which allows Matlantis™ to predict transition states and correctly represent unknown systems. 11

© Preferred Computational Chemistry, Inc. All Rights Reserved. Customer testimonial: Toyota Motor Corporation Geometry optimization of a Li-based battery material by Toyota Motor Corporation 29 times slower than Matlantis 69 times slower than Matlantis Matlantis™ DFT on a national supercomputer DFT on Toyota’s supercomputer x Highest speed “Previously, we had to search our database to find new materials. It took about three months. Matlantis does that in a week. With Matlantis, we can calculate roughly 10 million structures in a year. This can be quite an advantage to find promising candidates.” 12

13 Use cases

© Preferred Computational Chemistry, Inc. All Rights Reserved. Catalyst case: Renewable synthetic fuel catalysts https://matlantis.com/en/cases/calculation001/ Aim: Identify dopant species for Fischer-Tropsch catalysts (improve CO dissociation rate by optimizing catalyst promoters) Approach: Screen 9300 configurations with varying compositions (cobalt host metal) Result: Found effective dopant species that lower the activation energy by nearly 40% Calculation time: Two weeks (vs. years by DFT) MAE = 0.1 eV Comparison of activation energies DFT vs. Matlantis™ Activation energies of methanation reactions for synthesis gas on Co(0001) Effect of promoters on relative activation energy Matlantis CO dissociation (Co+V) C5+ formation scheme 🡪 Activation energy lowered by 55 kJ/mol – 40% 14

© Preferred Computational Chemistry, Inc. All Rights Reserved. Battery case: Li diffusion in solid electrolytes https://www.kek.jp/ja/newsroom/2016/06/22/1133/ Root mean square displacement (at 523 K) Diffusion coefficient Activation Energy（meV） Matlantis™ DFT [1] Exp. [2] 230 210 242 [1] Mo et al. Chem. Mater. (2012) 24, 15-17 [2] Y. Kato et al. Nat. Energy 1, 16030. https://matlantis.com/en/cases/calculation004/ Matlantis LGPS: well-known Li-ion conductor Ion diffusion Aim: Predict the diffusion coefficient of lithium ions in LGPS Approach: MD simulation in NVT ensemble Result: Reproduced the characteristic features of Li diffusion. The results are almost identical to that of DFT Calculation time: 2-3 days for the Arrhenius plot (vs. around a year by first-principles molecular dynamics) 15

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Objectives and Targets of the Study Dr Tasuku Onodera 16

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Mechanism and Base Oil Selection Guidelines Dr Tasuku Onodera Good Bad 17

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Base oil screening methods Step 2 Dr Tasuku Onodera 18

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Characterization of oil film structure (step 2) Dr Tasuku Onodera 19

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Characterization of oil film structure (step 2) Dr Tasuku Onodera 20

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Mapping and Rapid Screening Dr Tasuku Onodera 21 Teacher data Predicted data Available compounds

© Preferred Computational Chemistry, Inc. All Rights Reserved. Lubricant case: In silico screening of base oil molecular structures Proposal of base oils for better processing Dr Tasuku Onodera 22 Teacher data Predicted data Available compounds

© Preferred Computational Chemistry, Inc. All Rights Reserved. Matlantis™ > dramatically accelerates mechanisms elucidation > enables large-scale screening for fast chemicals and materials discovery > facilitates the digitalization of R&D departments Wrap-up Recent updates… o Contract research activities for non-expert customers o “Light-PFP” feature allowing users to study up to 1 000 000 atoms o 13 billion-parameter pre-trained Large Language Model (future co-pilot) https://matlantis.com 23

24 Visit us during ACHEMA 2024: Hall 11.0 Stand F77 Q&A time